Tool‑augmented language‑model agents rely on interface text at inference time—tool descriptions, argument schemas, and the list of candidate tools. Even when the underlying tool semantics stay unchanged, these interfaces often evolve after deployment, causing a reliability issue known as interface drift. We investigate whether a lightweight, training‑free repair layer can restore robustness without retraining the base model.\
Our proposed SchemaShield‑Lite consists of three components:\
- Schema validation: ensure the model's output conforms to the tool's JSON schema;\
- Candidate‑set retry: if the returned tool name is missing from the current candidate list, search the full candidate set again;\
- One‑shot constrained repair: upon detecting a mismatch, generate a corrected output via constrained sampling in a single step.\
Experiments on Qwen3.5‑9B show that SchemaShield‑Lite raises DICE‑300 accuracy from 0.717 to 0.813 and BFCL‑200 accuracy from 0.580 to 0.805, recovering 19/22 and 33/45 drift‑induced failures respectively, with no observed repair harms. Simple retry does not account for these gains. Ablation studies reveal that candidate‑set drift, especially the presence of distractor tools, is the dominant source of damage, while description or schema edits alone are less harmful.\
These findings position validation‑and‑repair as a practical robustness layer for deployed LLM‑agent tool usage.\
Review